Handwritten Text Segmentation in Scribbled Document via Unsupervised Domain Adaptation

Abstract

Supervised learning methods have shown promising results for the handwritten text segmentation in scribbled documents. However, many previous methods have handled the problem as a connected component analysis due to the extreme difficulty of pixel-level annotations. Although there is an approach to solve this problem by using synthetically generated data, the resultant model does not generalize well to real scribbled documents due to the domain gap between the real and synthetic dataset. To alleviate the problems, we propose an unsupervised domain adaptation strategy for the pixel-level handwritten text segmentation. This is accomplished by employing an adversarial discriminative model to align the source and target distribution in the feature space, incorporating entropy minimization loss to make the model discriminative even for the unlabeled target data. Experimental results show that the proposed method outperforms the baseline network both quantitatively and qualitatively. Specifically, the proposed adaptation strategy mitigates the domain shift problem very well, showing the improvement of baseline performance (IoU) from 64.617% to 85.642%.